license: mit
task_categories:
- video-classification
- text-classification
- zero-shot-classification
language:
- en
pretty_name: Ego2Exo
size_categories:
- 10K<n<100K
Dataset Preparation.
Our Ego2Exo benchmark is curated from the EgoExo4D dataset released by Meta. A complete documentation of the dataset preparation including selecting video segments, narrations and label set is available at this colab notebook. Note that you have to first download the annotations and keystep labels from the original dataset.
Dataset Download Instructions
The original video segments should be downloaded from the original EgoExo4D website from Meta by following their guidelines. You might have to first sign a license form to access and download the dataset, which might take upto 48 hours for getting approved. You can download the cli installer to enable faster downloads. Make sure that the installation works by running egoexo --help which should return the documentation guide.
In our paper, we only used the pre-extracted Omnivore features for all the video segments as input to the classification network. If you only want to download the pre-extracted features for the takes corresponding to our dataset, use the following command. Make sure to provide the argument data_dir or specify where the files should be downloaded.
sh omnivore_features_download.sh <data_dir>
You can also download the complete videos corresponding to our takes. The whole dataset is quite large, so it is recommended that you only download the videos used in our benchmark with the following command. Make sure to provide the argument data_dir or specify where the files should be downloaded.
sh takes_download.sh <data_dir>